面对高光谱影像分类的半监督阶梯网络

提出一种半监督阶梯网络用于对高光谱影像进行分类,以解决小样本条件下基于堆栈式自编码器的高光谱影像分类方法分类精度不高的问题。首先,该网络以堆栈式自编码器为基础,在编码器和解码器之间增加横向连接参数构建阶梯网络,以使网络适合半监督分类;然后将无监督损失函数与有监督损失函数之和作为最终优化的目标函数,采用半监督的方式对整个网络进行训练。为进一步提高分类精度,提取局部二值模式纹理特征进行分类实验。实验结果表明:提出的半监督阶梯网络能够较好地解决高光谱影像分类小样本问题;且LBP纹理特征能够有效提高分类精度。

Semi-Supervised Ladder Network for Hyperspectral Image Classification

A semi-supervised ladder network is proposed to classify hyperspectral images to solve the problem of low classification accuracy of hyperspectral image classification methods based on stacked autoencoder under the condition of limited samples.Firstly,the network based on the stacked autoencoder adds the skip connection parameters between the encoder and the decoder to construct a ladder network which makes the network suitable for semi-supervised classification.Then,the sum of unsupervised loss function and supervised loss function is used as the loss function of the final optimization.The network is trained using semi-supervised training method.In order to further improve the classification accuracy,the local binary model texture feature is used for the classification experiments.The experimental results demonstrate that the proposed semi-supervised ladder network can deal with the problem of small sample classification of hyperspectral image classification.In addition,the LBP texture feature can improve the classification accuracy effectively.